Exploring self-supervised learning in Multiview captcha recognition
Mukhtar Opeyemi Yusuf, Divya Srivastava, Riti Kushwaha · 2023
Text-based CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) systems play a crucial role in safeguarding online platforms against automated bot attacks. However, with the increasing complexity of text-CAPTCHA designs, traditional supervised learning approaches face challenges in accurately recognizing the underlying text in text-CAPTCHA images. One of the main challenges is the limited availability of labeled data, as manually annotating large amounts of text-CAPTCHAs is a time-consuming and expensive process. To address these challenges, we propose a novel Self-Supervised Multiview Captcha Recognition Model (SS-MCR) that leverages self-supervised learning with contrastive loss to learn representations from limited labeled data. Our model redefines the text-CAPTCHA solution as a multiview problem, using image manipulations to create a pseudo-view of the original single-view image. The SS-MCR employs a siamese architecture with 2D convolutional neural networks and recurrent networks to learn robust features from both the pseudo-view and the original view. The contrastive loss encourages the model to distinguish between corresponding pairs and non-corresponding pairs, facilitating the extraction of informative and discriminative features. Through extensive experiments on synthesized datasets and publicly available datasets, we demonstrate that the SS-MCR outperforms traditional supervised learning models and achieves state-of-the-art performance in text-CAPTCHA recognition. Our proposed approach offers a promising direction for improving text-CAPTCHA security while reducing the dependence on large labeled datasets, making it a valuable contribution to the field of computer vision and captcha recognition.